JMIR Publications

ABSTRACT

Background:

Retrieval-Augmented Generation (RAG) is increasingly applied in medicine and nursing to enhance large language models. However, there remains a lack of comprehensive understanding regarding its specific architecture and its application in medical reasoning.

Objective:

This study aims to investigate the current state and emerging trends of RAG in the medical and nursing domains.

Methods:

PubMed, Web of Science, IEEE Xplore, and arXiv were searched for relevant articles using queries that combine terms related to RAG, large language model, medical, and nursing. This review was conducted following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines.

Results:

A total of 917 articles were retrieved, of which 67 met the inclusion criteria. The RAG frameworks included in this review were categorized into five functional types: text-based RAG (n=36), knowledge graph-enhanced RAG (n=17), agentic RAG (n=6), multimodal RAG (n=2), and plug-and-play RAG (n=6). Based on a staged decomposition of the RAG workflow into intent recognition, knowledge retrieval, knowledge integration, and generation stage, we analyzed the specific techniques employed at each stage across the included studies. Despite the growing emphasis on reasoning, only 26 studies incorporated explicit reasoning mechanisms, and few aligned with the procedural logic of clinical or nursing workflows.

Conclusions:

This study revealed that recent advancements in medical and nursing RAG frameworks demonstrate four key transformations: from surface-level matching to contextualized intent recognition; from vague semantics to logic-driven dynamic retrieval; from passive to active knowledge retrieval; and from simple aggregation to coherent context construction. However, most RAG systems in medical and nursing have not yet introduced reasoning methods, and those that have are still predominantly reliant on statistical associations. This underscores the necessity of incorporating causal mechanisms to achieve more profound and domain-specific reasoning.